No headline act this week — the plumbing got serious — the agent stack hardened.

1. 01

   E2B sandboxes can now pause and resume with full memory state

   Sandbox persistence (public beta) snapshots filesystem and memory — pause an idle agent session, resume it ~1s later exactly where it left off, up to 30 days. Auto-pause replaces shutdown-on-timeout. If we ever run generated code in a sandbox, this is the primitive that makes per-user sessions cheap to keep warm.

2. 02

   Vercel Blob drops long-lived tokens — OIDC auth and signed URLs land

   Blob now supports OIDC authentication (no more long-lived tokens) and time-bound signed URLs for upload/download/delete. Plus elastic build machines auto-adjust to dodge OOM failures. Real hosting-layer hardening on the surface we deploy against.

3. 03

   Cloudflare hardens Workers — egress control, bulk secrets, Agent Skills

   Workers can now filter public egress through Gateway policies, manage up to 100 secrets per API call, and schedule Workflow instances via cron in the binding. The Agents SDK v0.14.0 adds an Agent Skills primitive — the skills-as-table-stakes arc reaches the edge runtime.

4. 04

   Anthropic trims the API edges — and says Claude writes 80% of its own code

   Smaller, real API wins: the advisor tool gains max_tokens, and refusals with no output stop being billed. Alongside, Anthropic published that over 80% of code merged at the company is now written by Claude — a data point worth sitting with.

E2B shipped sandbox persistence (public beta): you can pause a running sandbox and resume it later with not just the filesystem but the full memory state — running processes, loaded variables, everything — restored in about a second. Pausing costs ~4s per GB of RAM; sandboxes can stay paused up to 30 days. A new auto-pause mode replaces shutdown-on-timeout, so idle agent sessions hibernate instead of dying.

We don't run a sandbox today, but this is the exact primitive the agent-execution future needs. If Onlook ever executes user-generated code per project, persistence is what makes a per-user session cheap to keep warm between interactions — resume in a second rather than cold-booting and replaying state. It pairs directly with the Anthropic self-hosted-sandboxes thread we've been tracking.

Vercel Blob picked up two real security upgrades: OIDC authentication, which eliminates long-lived storage tokens, and time-bound signed URLs that scope upload, download, inspection, and deletion to a window. Separately, elastic build machines now auto-adjust resources to prevent out-of-memory failures mid-deploy.

All three land on the hosting surface we actually deploy against. The Blob auth changes are the kind of thing that quietly removes a whole class of credential-leak risk; the OOM-protected builds remove a flaky-deploy failure mode. Nothing flashy, all directly useful.

A cluster of Workers hardening this window: you can now filter a Worker's public-internet egress through Gateway policies, manage up to 100 secrets in a single API call, and attach cron schedules to Workflow instances directly in the binding (no separate handler). The Agents SDK v0.14.0 also added an Agent Skills primitive — the same "skills" surface that keeps showing up across the category, now at the edge.

One migration note: the pipeline binding config field was renamed to stream (old field deprecated). Egress control and bulk secrets are the items that matter if we ever run agent code on Workers — they're exactly the containment knobs you want.

Two quiet but real API changes on the Claude Platform: the advisor tool gained a max_tokens cap (lower latency and cost when you don't need a full advisor response), and requests that return a refusal with no output are no longer billed. Both are the kind of edge-trimming that adds up across a high-volume agent loop.

The line worth sitting with came from Anthropic itself: over 80% of the code merged at the company is now written by Claude. Whatever the exact accounting, it's the clearest public signal yet of where agent-written production code is heading — and we're building a tool squarely in that current.

A

Alex Albert

@alexalbert__ · Jun 4

Published internal data: over 80% of all code merged at Anthropic is now written by Claude.